Assessment of fetal maturation age by heart rate variability measures using random forest methodology
Introduction
Since heart rate patterns are one of the few signals obtainable from the fetus easily and non-invasively, they are predestinated for assessing the maturating fetal autonomic control and its disturbances [1], [2], [3], [4], [5]. The complex behavior of the maturating fetal autonomic nervous system (ANS) was recently described by means of corresponding heart rate variability (HRV) indices. However, the maturation itself is a complex and non-linear process. Interpretation of non-linear complexity characteristics is ambiguous and has been discussed with the formation of fetal behavioral states [6], [7], [8]. On the basis of changes in the power spectra, van Leeuwen et al. showed that maturation is characterized by non-linear characteristics [9] and different stages of fetal development were discussed in association with the increasing influence of the different branches of the ANS [3], [10]. However, to date the complex maturation process was mainly approximated by linear characteristic curves in univariate regression models using linear and non-linear HRV indices [6], [9], [10], [11]. The investigation of the performance of the best predicting HRV indices in a multivariate model using non-linear and complex maturation characteristic curves is pending. Hoyer et al. recently proposed a fetal autonomic brain age score (fABAS) for the assessment of fetal age, based on MLR models according to universal developmental characteristics [12], [13]. However, parameter selection for MLR models in these investigations was based on a pre-selection of HRV indices. Since even the maturation characteristic curves are not known and they can be different for different HRV indices, a complex non-linear approach without pre-setting is required.
With regard to that, classification and regression tree (CART) methodology such as Random Forest (RF) may provide potential advantages [14]. The non-linear and complex data structures of RF provide an ambitious technique for data mining, e.g. in geoinformatics [15], [16] or computational biology [17], [18], [19]. A direct comparison between ordinary linear regression and RF was done by Nir et al., who investigated the performance of both models for the assessment of the nociception level under anesthesia [20]. They found that the advanced, non-linear approach performed better than linear regression. However, a general statement on the superiority of RF compared to other linear and non-linear models is limited due to the fact that the accuracy of predictions is biased by different methodological approaches and results heavily depend on the underlying problem [21]. With respect to clinical applications, researchers often try to obtain as much information as possible from the investigated process. Modeling such processes with common linear methods requires previous knowledge of interactions between variables and explicit modeling of non-linearities. RF as a non-linear, multivariate regression and classification methodology is able to overcome this problem, even when values are missing [14]. Additionally, the assessment of the parameter importance employed in RF provides a beneficial tool for the identification of important variables avoiding pre-selection of HRV parameters on the basis of single characteristic curves. Single applications of RF to heart rate time series in adults have been reported with satisfying results, e.g. in connection with sleep state classification [22], classification of cardiac rhythms [23], risk stratification for arrhythmic cardiac death [24], or the prediction of cardiovascular and cerebrovascular events [25]. As far as is known to the authors, the only application of RF in the context of fetal development was published by Peterek et al. who classified pathological, suspect and normal fetal states based on cardiotocography (CTG) measurements [26]. However, due to its restricted temporal resolution, CTG is of limited appropriateness for the precise assessment of fast heart rate modulation [27]. According to the “developmental origins of adult disease (Barker) hypothesis” (also known as fetal programming) [28], the precise evaluation of the normal development is important with respect to the early identification of fetal developmental disorders since these have implications for health problems in later life which cannot completely be compensated for by later postnatal therapies [29].
The objective of the present work is to evaluate the capability of a complex maturation model, using previously developed HRV indices obtained from high resolution fetal magnetocardiographic (fMCG) recordings and complex maturation functions to predict fetal gestational age (GA) for the assessment of normal fetal development. Further, we compare results of GA prediction from RF methodology with linear, multivariate age regression.
Section snippets
Subjects
359 fMCG recordings were taken from healthy, singleton fetuses with an age between 21–41 weeks of gestational age (WGA). Recordings from subjects with intrauterine growth restriction, non-reassuring non-stress test based on conventional CTG, known chromosomal abnormalities or congenital abnormalities based on ultrasound diagnosis, fetal arrhythmia or previous exposure to synthetic steroids in utero were not considered for analysis. Additionally, maternal exclusion criteria were: administration
Results
We determined as the optimal choice for the number of trees grown in RF, providing sufficient model complexity at a moderate calculation time. Further increase of ntree showed only negligible effects on the OOBerror (See Fig. 1).
After defining the model parameters ntree and mtry, fetal age regression was performed by RF for fetal states HRP I, HRP II and state independent measurements (30 min). RF showed a slightly better prediction accuracy in all three groups compared to MLR (HRP I:
Discussion
In this work we presented a novel approach for determining fetal maturation considering non-linear characteristics. In order to consider multiple interactions between HRV indices, we used a parameter set without any preselection on the basis of single characteristic curves of HRV parameters. As far as known to the authors, this is the first time that such a large number of linear and non-linear HRV parameters was included in a non-linear, multivariate analysis of fetal heart rate time series,
Conclusion
Within this work, we presented a novel approach for the assessment of fetal maturation, considering complex, non-linear characteristic curves. Relating to this, RF provides advantageous characteristics and is appropriate for the evaluation of fetal maturation age. Our results support the importance of complexity measurements and their relationships across multiple scales. Those inter-relationships may help to gain a deeper insight into the development of complex coordination mediated by the ANS.
Summary
Prenatal risk factors can permanently change the fetal brain development and cause diseases in later age. However, prenatal functional diagnosis is limited and requires innovative concepts. What is required, is a more sophisticated analysis of the fetal neuro-vegetative (autonomic) control which provides valuable information. Since heart rate is one of the few signals that can be obtained non-invasively from the fetus, HRV analysis is uniquely suited to assess fetal functional autonomic brain
Confilct of interest statement
None Declared
Acknowledgements
We thank Franziska Jaenicke, Esther Heinicke, Anja Rudolph, Ulrike Wallwitz, Isabelle Kynass, Franziska Bode, Janine Tegmeyer, Kathrin Kumm, and Liviu Moraru (Team of the Biomagnetic Center/Hans Berger Department of Neurology and the Department of Obstetrics, Jena University Hospital) for their contributions to fMCG recordings and heart beat trigger preprocessing.
References (46)
- et al.
Fetal heart rate variability and complexity in the course of pregnancy
Early Hum. Dev.
(1999) - et al.
Indices of fetal development derived from heart rate patterns
Early Hum. Dev.
(2009) - et al.
Linear and nonlinear analysis of heart rate patterns associated with fetal behavioral states in the antepartum period
Early Hu. Dev.
(2007) - et al.
Random forests for land cover classification
Pattern Recognit. Lett.
(2006) - et al.
Random forests as a tool for ecohydrological distribution modelling
Ecol. Model.
(2007) - et al.
Random forests-based differential analysis of gene sets for gene expression data
Gene
(2013) - et al.
A comparison of random forest regression and multiple linear regression for prediction in neuroscience
J. Neurosci. Methods
(2013) - et al.
Sleep stages classification based on heart rate variability and random forest
Biomed. Signal Process. Control
(2013) - et al.
A computer system for the numerical analysis of nonstress tests
Am. J. Obstet. Gynecol.
(2002) - et al.
Fetal development assessed by heart rate patterns-time scales of complex autonomic control
Comput. Biol. Med.
(2012)
Heart rate variability in the individual fetus
Auton. Neurosci.Basic Clin.
Assessment of fetal neurodevelopment via fetal magnetocardiography
Exp. Neurol.
Fetal heart rate variability reveals differential dynamics in the intrauterine development of the sympathetic and parasympathetic branches of the autonomic nervous system
Physiol. Meas.
Complexity analysis of the fetal heart rate variabilityearly identification of severe intrauterine growth-restricted fetuses
Med. Biol. Eng. Comput.
An estimate of fetal autonomic state by time-frequency analysis of fetal heart rate variability
J. Appl. Physiol.
Fetal development of complex autonomic control evaluated from multiscale heart rate patterns
Am. J. Physiol. Regul. Integr. Comp. Physiol.
Changes in the frequency power spectrum of fetal heart rate in the course of pregnancy
Prenat. Diagn.
Development of integrative autonomic nervous system functionan investigation based on time correlation in fetal heart rate patterns
J. Perinat. Med.
Development of multiscale complexity and multifractality of fetal heart rate variability
Auton. Neurosci.Basic Clin.
Fetal functional brain age assessed from universal developmental indices obtained from neuro-vegetative activity patterns
PloS one
Fetal autonomic brain age scores, segmented heart rate variability analysis, and traditional short term variability
Front. Hum. Neurosci.
Random forests
Mach Learn.
Cited by (17)
PSO optimized 1-D CNN-SVM architecture for real-time detection and classification applications
2019, Computers in Biology and MedicineCitation Excerpt :Principal Component Analysis (PCA), Singular Value Decomposition algorithm (SVD) and Linear Discriminant Analysis (LDA) are some of the frequently used feature extraction techniques [16,17]. The major classification algorithms are Decision Tree (DT), K-Nearest Neighbor (KNN), Perceptron, Random Forest and Support Vector Machine [18–20]. The computational time and complexity are high in the traditional methods as they use separate techniques for feature extraction and classification, and therefore they cannot give better performance in real-time applications.
Hybrid-FHR: a multi-modal AI approach for automated fetal acidosis diagnosis
2024, BMC Medical Informatics and Decision MakingHierarchical Attentive Network for Gestational Age Estimation in Low-Resource Settings
2023, IEEE Journal of Biomedical and Health InformaticsPredicting the Mortality of ICU Patients by Topic Model with Machine-Learning Techniques
2022, Healthcare (Switzerland)Non-linear Methods Predominant in Fetal Heart Rate Analysis: A Systematic Review
2021, Frontiers in MedicineStudy on the Impact of Online Courses for Pregnant and Lying-In Women on Maternal and Infant Health during the Epidemic
2021, Journal of Healthcare Engineering